Anthropic reports that Claude agents helped identify a bacteriophage system it calls array-associated reverse transcriptases (ART), linking a reverse-transcriptase gene with an accessory gene and repeated non-coding DNA. The company says early human-run experiments found distinct short RNAs from the array. That is evidence of RNA expression, not proof of the enzyme’s function, a gene-editing capability or a medical application. A Nature update also reports a separate research group’s prior work on the system, leaving discovery history and data provenance questions to resolve.
What Anthropic says Claude found
Anthropic’s September 23, 2026 announcement describes array-associated reverse transcriptases, or ARTs: a reverse-transcriptase gene next to a partner gene and an array of evenly spaced, non-coding DNA repeats, found mainly in bacteriophages (viruses that infect bacteria). The underlying reverse transcriptase was already known from earlier studies; Anthropic’s specific claim is that its agents recognized the surrounding repeat-and-partner pattern as a possible system. The arrangement resembles CRISPR arrays at a structural level. Resemblance is a useful clue for researchers, not evidence that ART works like CRISPR.
The agent scale is notable; the number is not a quality score
Anthropic says roughly 950 Claude agent sessions searched for 21 hours and used about 210 million tokens. The company reports that the workflow gathered more than 200,000 reverse transcriptases, proposed 3,500 candidate systems, and narrowed those to 20 reports for closer review. Those figures describe the company’s process, not a head-to-head benchmark or a measure of accuracy. The scientific value lies in whether the candidates survive careful review and experiments—not in how many agents ran or tokens they consumed.
What the wet-lab work does—and does not—show
Anthropic reports that initial experiments detected the repeat array as distinct short RNAs. That is an early experimental observation supporting RNA expression and processing. Anthropic also says its human scientists performed all laboratory work; the model helped search sequence data, compare candidates, review literature and prepare reports. The announcement says the system’s primary function is still unknown and that follow-up experiments are underway. The preprint is an early report, so its claims should be treated as provisional until methods, interpretation and results receive wider scrutiny.
A discovery-history question deserves careful reporting
Nature’s September 25 report was later updated with a separate scientist’s account: Mario Rodríguez Mestre of the University of Copenhagen said his group had studied the microbial system for years and that he had uploaded information about it to Anthropic’s models. Nature reported Anthropic’s response that the model responsible was not trained on user transcripts. These statements do not, by themselves, settle what the agents accessed, how the pattern was independently recognized, or how the system’s history should be credited. They do make data lineage, prior-art searches and transparent attribution central questions for readers to follow. We are reporting the two accounts without treating the disagreement as resolved.
Why this is not a new CRISPR or a treatment
“CRISPR-like” in this story refers to the visual organization of repeated DNA sequences and the possibility that those sequences have a role in a biological system. Anthropic has not shown that ART edits genomes, cuts or copies DNA in a programmable way, or can be used in a therapy. It has not announced a medicine, clinical study or validated biotechnology product. A candidate mechanism needs biochemical characterization, functional experiments, independent replication and safety assessment before any translational claim would be justified.
What would make the result stronger next
The next meaningful milestones are concrete: clarify the system’s components and activity; test competing explanations for the short RNAs; publish methods and data sufficient for outside groups to inspect the claim; establish how the prior work and training or input data relate to the agent’s output; and see whether independent researchers reproduce the findings. This is also the useful lens for AI in biology more broadly: agents may help sift enormous datasets and produce hypotheses, while accountable scientists remain responsible for experimental design, evidence, interpretation and credit. For a different kind of multi-step AI workflow, see our guide to stateful agents versus simple retrieval.
Frequently asked questions
What does ART mean in Anthropic’s wet-lab announcement?
ART means array-associated reverse transcriptases: a reported arrangement of a reverse-transcriptase gene, a neighboring partner gene and an array of non-coding DNA repeats, found mainly in bacteriophages.
Did Claude perform the wet-lab experiments?
No. Anthropic says human scientists performed all laboratory work. Claude agents searched and analyzed sequence data, reviewed literature and helped identify candidates for human follow-up.
What did the experiments confirm?
Anthropic reports early evidence that the repeat array is expressed as distinct short RNAs. The primary biological function and enzyme mechanism remain unknown.
Is ART a new CRISPR gene-editing tool?
No demonstrated gene-editing function has been reported. The CRISPR comparison describes a resemblance in repeat-array structure; it does not establish that ART edits DNA or has therapeutic uses.
How many AI agents and tokens did Anthropic report?
Anthropic says roughly 950 agents searched for 21 hours and used about 210 million tokens. These are company-reported workflow figures, not an independent accuracy benchmark.
Was the ART system previously known?
Anthropic says the underlying reverse transcriptase had been identified earlier, while the associated array and partner-gene pattern was the notable observation. Nature later reported a researcher’s account that his group had studied the system for years; the discovery history and data lineage remain questions for further clarification.
Sources & further reading
Primary sources checked Oct 1, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Claude discovers a novel enzyme system with CRISPR-like repeats ↗Anthropic · Sep 23, 2026
- 2Anthropic’s AI biolab finds “CRISPR-like” DNA in viruses. What’s next? ↗Nature · Sep 25, 2026
- 3Technical preprint on AI-agent-assisted identification of array-associated reverse transcriptases ↗Anthropic research team · bioRxiv · Sep 22, 2026
Help us keep this useful. Send a correction or a primary source →
